Inverse Quantization Method Of Grayscale Image Based On Residual Network

With the vigorous development of the Internet of Things and the commercialization of virtual reality technology,it is often necessary to perform three-dimensional representation of the objective world in the computer.And 3D reconstruction is an interdisciplinary subject,involving advanced mathematics,computer vision,computer graphics and so on.The specific practical operation is to use sensors such as cameras to photograph real-life objects and scenes and process them with computer vision technology to obtain a complete 3D model of the object.Now with the development of various advanced scanning equipment,people can now obtain high-precision data from the surface of three-dimensional objects in a short time.Now mainstream technologies include 3D point cloud,multi-view stereo geometry,mesh reconstruction and optimization,texture mapping,Markov random field,image dequantization and so on.3D reconstruction is currently one of the core technologies in fields such as augmented reality(AR),mixed reality(MR),robot navigation,and autonomous driving.The high-resolution reconstruction technique for images refers to the use of compressed and noisy degraded images or video sequences frames to generate high-quality,high-resolution images,automatically filling in missing details,improving texture details,and reducing image noise.In existing image restoration models,whether based on CNN or Transformer structures,they often require a large amount of redundant training,which consumes significant computing resources.To address these issues,this paper explores network models based on deep learning for contour line repair and degraded image restoration,and propose corresponding improvement methods.The main work is as follows:(1)A residual network based geometric contour inpainting network is proposed.The error contour problem introduced after converting the grayscale image into a height map is transformed from a geometric point of view into a constrained optimization problem,and the residual mapping function is learned in a data-driven manner.The network also adopts a structural design based on the residual network and a normal-based loss function.Finally,the experiment proves that the 3D model generated by using the restored height map of this method is better than the effect of directly using the grayscale image for conversion,and the network also has better stability and robustness.(2)A Transformer-based image restoration network is proposed,which takes degraded grayscale images as input,such as lossy images,noisy or compressed images,and can generate high-quality grayscale images with rich details.The network structure consists of a shallow feature extraction module,a deep feature extraction module and a high-definition image reconstruction module.First,the shallow feature extraction module is employed to extract the low-frequency information from the image,To prevent information loss,a residual connection is directly transmitted to the final feature extraction module.Then the deep feature extraction module extracts the high-frequency information features of the image,and finally performs information aggregation in the image reconstruction module to restore high-quality images.In the final comparative experiment,it also proves the simplicity and efficiency of the network model.